It's correlated with the ability to solve logic puzzles.

It's also interesting because it's very very hard for base LLMs, even if you try to "cheat" by training on millions of ARC-like problems. Reasoning LLMs show genuine improvement on this type of problem.